Evaluate projections

Create a ZADU runner with your measures and original data, then pass a projection to measure(). Reuse the runner for projections of the same samples. The examples below use the arrays from the quickstart.

Configure a runner

from zadu import ExecutionConfig, ZADU

specs = [
    {"id": "tnc", "params": {"k": 10}},
    {"id": "stress"},
]
runner = ZADU(specs, original, execution=ExecutionConfig(memory_budget="512MiB"))
scores = runner.measure(projection)

original has shape (n, d) and projection has shape (n, p), with the same samples in the same row order. The runner stores a read-only copy of original; later changes to your array do not alter that runner. Create a new runner to evaluate a different original dataset.

The API reference lists all constructor arguments and return formats. Use return_local=True for pointwise scores.

Inspect execution

scores contains the measure results. last_run_info describes the most recent completed call:

info = runner.last_run_info
print(info["backend"])
print(info["planned_peak_bytes"])
for resource in info["resources"]:
    print(resource["kind"], resource["provider"], resource["reused"])

Use provider to see where each calculation ran and reused to see whether a cached result was used. planned_peak_bytes estimates the memory covered by the execution budget; it excludes the original input copy and is not a process memory measurement.

Spherical coordinates

For original longitude/latitude coordinates in radians, geodesic=True uses angular distances on a unit sphere when building shared original-space distance and neighbor resources. Projection distances remain Euclidean.

import numpy as np

# Columns are longitude, latitude; convert degree input to radians.
coordinates = np.deg2rad([[0, 0], [10, 0], [0, 10], [10, 10]])
map_projection = np.array([[0, 0], [1, 0], [0, 1], [1, 1]], dtype=float)
runner = ZADU([{"id": "tnc", "params": {"k": 1}}], coordinates, geodesic=True)
scores = runner.measure(map_projection)

This option does not transform the coordinates or turn coordinate-based calculations such as Procrustes alignment, CADI angles, or Gap Index areas into spherical versions of those measures. Restrict a spherical evaluation to measures whose use of the shared distances matches your intended definition. Accelerator providers use NumPy for geodesic resources.

Direct calls

For a single calculation, you can also call a measure function directly. Direct calls do not reuse a runner's caches or apply its execution budget.